Emotion Classification in Western Languages

With the theory of Natural Semantic Metalanguage (NSM), Wierzbicka (1992) was the first to propose a series of semantic primitives such as I, you, someone, this, want, don’t want,

think, say, imagine, feel, know, good, and bad, and tried to decompose emotion into complex

events involving a cause and a mental state in terms of simple and non-technical terms. The following was her exemplification of the emotion Disappointment:

X feels something; sometimes a person thinks something like this: (1) something good will happen; (2) I want this; after this, this person thinks something like this: I know

10

now: this will not happen; because of this, this person feels something bad. X feels like this (p. 548).

Using these prototypical scripts or scenarios in terms of wants, thoughts, and feelings, even the basic emotion can be rigorously and revealingly portrayed. Furthermore, the differences between apparent synonyms like unhappy and sad can be fully specified according to their conceptual structures, which may first seem “fuzzy” or overlapping. However, the main purpose of NSM is not to classify emotion words but to emphasize the specification of each emotion concept and thus leads to more research on the lexical semantics of emotion words.

For example, Semin and Fielder (1991)’s Linguistic Category Model (LCM) distinguished action and stative verbs by several semantic criteria, one of which being positive or negative valence, a hedonic index to emotions. In LCM, verbs are divided into state verbs (e.g., to love, to admire), state action verbs (e.g., to surprise, to amaze), interpretative action verbs (e.g., to cheat, to help), and descriptive action verbs (e.g., to kick, to kiss). While state verbs are used to describe mental and emotional states, state action verbs often express emotion consequences of an action with a beginning and end. Together with interpretative action verbs, which often refer to a general class of behaviors or actions, these three types of verbs are all associated with semantic valence connotations. Although whether descriptive action verbs, which are distinguished from interpretative action verbs by invariant physical features of event types, could be related to affects is still under discussion, Semin and Fielder (1991)’s

11

model brings the necessity to incorporate the lexical analysis of words to the encoding of our emotion experiences.

It is worth noting that the LCM was based on one general verb taxonomy between states and events (Davidson, 1971; Dowty, 1979; Rappaport Hovav & Levin, 1998). Stative verbs describe an enduring situation or a state of being, while eventive verbs entail a series of causal changes throughout the process in event structure (Dowty, 1979), the representation of events and their participants. Eventive verbs can also be differentiated from stative verbs by their entailment of the conceptual units involving CAUSE, BECOME, CHANGE and resulting STATE, while stative verbs do not have this causal chain, which contributes to different lexical complexity to the two types of verbs. The distinction between eventive and stative verbs has psychological reality, which was evidenced in Gennari and Poeppel’s (2003) study. With lexical decision task (LDT), subjects had longer reaction time in processing eventitve verbs compared with stative ones (e.g. vanish vs. exist). The authors concluded that this result cannot be explained by thematic roles or argument structures the verb meanings carry, but should be attributed to different event structure properties activated during processing.

Apart from lexical distinctions, some scholars focus more on the semantic and syntactic nature of emotion words. Levin (1993) classified emotion verbs based on the transitivity and the experiencer’s syntactic position in a sentence, labeling them as “Psych-verbs” with four

12

subcategories. Amuse and Admire Verbs are transitive with object and subject being experiencers respectively, whereas Appeal and Marvel Verbs are intransitive with experiencers as object and subject respectively. Goy (2000) also analyzed Italian adjectives using a frame-based approach integrating case-frame semantics, the generative lexicon and the prototype theories. She observed that there are three interpretations/readings of emotion adjectives: stative (e.g. cheerful boy), manifestative (e.g. affectionate letter) and causative (e.g. amusing movie), and classified Italian emotion adjectives based on the explanation from qualia structures (Pustejovsky, 1995), the specification of how a lexicon’s arguments and events are connected to modify relations in the semantic composition within the affective noun phrases. According to her proposal, the semantic representation of emotion states can be characterized as knowledge structures that also encode prototypical sequences of actions and events, which are very similar to scripts containing the beliefs, the reactions, the causes, and the consequences of the emotion state (Fehr & Russell, 1984; Shaver et al., 1987;

Wierzbicka, 1992). However, these linguistic approaches to the classification of emotion are often restricted to specific lexical categories, such as verbs or adjectives, and hence cannot provide a comprehensive model to characterize the emotion connotation across different lexical categories.

In contrast to traditional lexical and semantic views, the corpus-based analysis of linguistic expressions of emotion focuses on establishing norms for affective lexicon and

13

detecting the structures in which emotion concepts may be embedded. Affective Norms for English words (ANEW), WordNet Affect, and Berlin Affective Word List Reloaded (BAWL-R) are the most used materials in research of visual word processing in western languages. In ANEW and BAWL-R, a set of English and German affective words was manually rated on the valence, arousal and dominance/imageability; however, only BAWL-R assigned affective words to three different word classes: nouns, verbs, and adjectives. Unlike the previous resource of affective lexicon, WordNet Affect is not only an extension of WordNet domain to label affective concepts to a subset of synonymous sets (so-called

synsets) but also labels a semantic domain to provide conceptual relations among word

senses used to group words hierarchically in WordNet. In particular, words in WordNet Affect are categorized into two groups: direct emotion words that denoting emotion states (e.g. happy, fear), and indirect emotion words that eliciting emotions (e.g. snake, monster).

Although the interpretations vary among individuals for indirect emotion words due to different causality they evoke, the general affectivity is considered collective imagination and calculated by semantic affinity with affective lexical concepts (Strapparava et al., 2006). In addition to valence tagging, WordNet Affect also annotated stative/causal interpretations of adjectives, a classification similar to Goy’s (2000) analysis. Unfortunately, most previous psycholinguistic experiments used affective words from ANEW and did not tell the (in-) direct emotion words apart, and this could yield qualitatively different results in terms of

14

processing and information encoding (e.g. different costs of episodic affective memory retrieval).